Damage simulations are essential for assessing structural integrity and understanding material failure under various conditions. However, their high computational cost poses challenges, particularly in applications requiring repeated simulations such as design optimization, uncertainty quantification, and real-time decision-making. Traditional model reduction methods like proper orthogonal decomposition and hyper-reduction can reduce simulation time but often require intrusive code modifications and are usually confined to classical finite element frameworks. In this study, an autoencoder-based non-intrusive model reduction approach for gradient-extended damage simulations [1] is introduced and investigated. Autoencoders are special neural networks that can be used to create a compact latent space representation of high-fidelity simulation data. After appropriate training, the autoencoder becomes capable of capturing complex nonlinear relationships within the data and allows for a significant reduction in computation time while maintaining high accuracy [2]. Furthermore, real-world data can be easily incorporated directly into the reduced-order model. The methodology is validated on several structural example problems. Different neural network structures and techniques are investigated for their impact on accuracy. The results demonstrate that the framework can generally achieve order-of-magnitude reductions in computational time while maintaining high accuracy in predicting damage evolution and structural behavior. Due to its robustness and efficiency, the approach is promising for applications requiring rapid simulations, such as real-time monitoring, predictive forecasting, and uncertainty quantification in, e.g., digital twins. [1] Brepols T., Wulfinghoff S., Reese, S., A gradient-extended two-surface damage-plasticity model for large deformations, International Journal of Plasticity, Vol. 129, 102635, 2020. [2] Simpson, T., Dervilis, N., Chatzi, E., Machine learning approach to model order reduction of nonlinear systems via autoencoder and LSTM networks, Journal of Engineering Mechanics, Vol. 147 (10), 04021061, 2021.